How does the the T-Net achieve translation invariance?

[ywangwxd]

I understand the t-net can estimate a 3x3 matrix for affine transformation, but 3x3 transformation is not translation invariant.
I am wondering how t-net achieve this.